Towards Interaction-level Video Action Understanding

Abstract

A huge amount of videos have been created, spread, and viewed daily. Among these massive videos, the actions and activities of humans account for a large part. We desire machines to understand human actions in videos as this is essential to various applications, including but not limited to autonomous driving cars, security systems, human-robot interactions and healthcare. Towards real intelligent system that is able to interact with humans, video understanding must go beyond simply answering ``what is the action in the video", but be more aware of what those actions mean to humans and be more in line with human thinking, which we call interactive-level action understanding. This thesis identifies three main challenges to approaching interactive-level video action understanding: 1) understanding actions given human consensus; 2) understanding actions based on specific human rules; 3) directly understanding actions in videos via human natural language. For the first challenge, we select video summary as a representative task that aims to select informative frames to retain high-level information based on human annotators' experience. Through self-attention architecture and meta-learning, which jointly process dual representations of visual and sequential information for video summarization, the proposed model is capable of understanding video from human consensus (e.g., how humans think which parts of an action sequence are essential). For the second challenge, our works on action quality assessment utilize transformer decoders to parse the input action into several sub-actions and assess the more fine-grained qualities of the given action, yielding the capability of action understanding given specific human rules. (e.g., how well a diving action performs, how well a robot performs surgery) The third key idea explored in this thesis is to use graph neural networks in an adversarial fashion to understand actions through natural language. We demonstrate the utility of this technique for the video captioning task, which takes an action video as input, outputs natural language, and yields state-of-the-art performance. It can be concluded that the research directions and methods introduced in this thesis provide fundamental components toward interactive-level action understanding.